Merge branch 'clang-format' of https://github.com/gangliao/Paddle; branch 'develop' of https://github.com/PaddlePaddle/Paddle into clang-format

CMakeLists.txt

@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)

cmake/cpplint.cmake

@@ -56,11 +56,14 @@ macro(add_style_check_target TARGET_NAME)
                 # cpplint code style
                 get_filename_component(base_filename ${filename} NAME)
                 set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
-                add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
+                add_custom_command(OUTPUT ${CUR_GEN} PRE_BUILD
                     COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
                             "--filter=${STYLE_FILTER}"
                             "--write-success=${CUR_GEN}" ${filename}
+                    DEPENDS ${filename} ${PROJ_ROOT}/paddle/scripts/cpplint.py
                     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+                add_custom_target(${base_filename}.cpplint DEPENDS ${CUR_GEN})
+                add_dependencies(${TARGET_NAME} ${base_filename}.cpplint)
             endif()
         endforeach()
     endif()

cmake/generic.cmake

@@ -187,7 +187,13 @@ function(cc_library TARGET_NAME)
     endif()
     
     # cpplint code style
-    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
+    foreach(source_file ${cc_library_SRCS})
+      string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+      if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+      endif()
+    endforeach()
+    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
 
   else(cc_library_SRCS)
     if (cc_library_DEPS)
@@ -239,6 +245,14 @@ function(nv_library TARGET_NAME)
         add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
         target_link_libraries(${TARGET_NAME} ${nv_library_DEPS})
       endif()
+      # cpplint code style
+      foreach(source_file ${nv_library_SRCS})
+        string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+        if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+          list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        endif()
+      endforeach()
+      add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS})
     else(nv_library_SRCS)
       if (nv_library_DEPS)
         merge_static_libs(${TARGET_NAME} ${nv_library_DEPS})

cmake/util.cmake

@@ -118,7 +118,6 @@ endfunction()
 macro(add_unittest_without_exec TARGET_NAME)
     add_executable(${TARGET_NAME} ${ARGN})
     link_paddle_test(${TARGET_NAME})
-    add_style_check_target(${TARGET_NAME} ${ARGN})
 endmacro()
 
 # add_unittest

paddle/framework/CMakeLists.txt

@@ -12,13 +12,15 @@ cc_test(variable_test SRCS variable_test.cc)
 cc_library(scope SRCS scope.cc)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-proto_library(attr_type SRCS attr_type.proto)
+proto_library(attribute_proto SRCS attribute.proto)
-proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
+proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto)
-proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
+proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto)
 cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
 
-cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope)
+cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto)
+
+cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator)
@@ -26,7 +28,7 @@ cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
 
-py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
+py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)

paddle/framework/attribute.cc

@@ -0,0 +1,85 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/framework/attribute.h"
+
+#include <vector>
+
+namespace paddle {
+namespace framework {
+
+template <>
+AttrType AttrTypeID<int>() {
+  return INT;
+}
+template <>
+AttrType AttrTypeID<float>() {
+  return FLOAT;
+}
+template <>
+AttrType AttrTypeID<std::string>() {
+  return STRING;
+}
+template <>
+AttrType AttrTypeID<std::vector<int>>() {
+  return INTS;
+}
+template <>
+AttrType AttrTypeID<std::vector<float>>() {
+  return FLOATS;
+}
+template <>
+AttrType AttrTypeID<std::vector<std::string>>() {
+  return STRINGS;
+}
+
+Attribute GetAttrValue(const AttrDesc& attr_desc) {
+  switch (attr_desc.type()) {
+    case paddle::framework::AttrType::INT: {
+      return attr_desc.i();
+    }
+    case paddle::framework::AttrType::FLOAT: {
+      return attr_desc.f();
+    }
+    case paddle::framework::AttrType::STRING: {
+      return attr_desc.s();
+    }
+    case paddle::framework::AttrType::INTS: {
+      std::vector<int> val(attr_desc.ints_size());
+      for (int i = 0; i < attr_desc.ints_size(); ++i) {
+        val[i] = attr_desc.ints(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::FLOATS: {
+      std::vector<float> val(attr_desc.floats_size());
+      for (int i = 0; i < attr_desc.floats_size(); ++i) {
+        val[i] = attr_desc.floats(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::STRINGS: {
+      std::vector<std::string> val(attr_desc.strings_size());
+      for (int i = 0; i < attr_desc.strings_size(); ++i) {
+        val[i] = attr_desc.strings(i);
+      }
+      return val;
+    }
+  }
+  PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
+  return boost::blank();
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/attribute.h

@@ -1,3 +1,17 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
 
 #include <boost/variant.hpp>
@@ -6,6 +20,9 @@
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+
+#include "paddle/framework/attribute.pb.h"
+#include "paddle/framework/op_desc.pb.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -14,13 +31,19 @@ namespace framework {
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                        std::vector<float>, std::vector<std::string>>
     Attribute;
+
 typedef std::unordered_map<std::string, Attribute> AttributeMap;
 
+template <typename T>
+AttrType AttrTypeID();
+
+Attribute GetAttrValue(const AttrDesc& attr_desc);
+
 // check whether a value(attribute) fit a certain limit
 template <typename T>
 class LargerThanChecker {
  public:
-  LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
     PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
   }
@@ -35,7 +58,8 @@ class LargerThanChecker {
 template <typename T>
 class DefaultValueSetter {
  public:
-  DefaultValueSetter(T default_value) : default_value_(default_value) {}
+  explicit DefaultValueSetter(T default_value)
+      : default_value_(default_value) {}
   void operator()(T& value) const { value = default_value_; }
 
  private:
@@ -78,7 +102,8 @@ class TypedAttrChecker {
   typedef std::function<void(T&)> ValueChecker;
 
  public:
-  TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
+  explicit TypedAttrChecker(const std::string& attr_name)
+      : attr_name_(attr_name) {}
 
   TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
     value_checkers_.push_back(EnumInContainer<T>(range));

paddle/framework/backward.cc

@@ -59,19 +59,17 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   //  If all input gradients of forwarding operator do not need to calculate,
   //  just return an NOP. Not return null ptr because NOP does not take
   //  too much time for calculation, but it is useful for simplifying logic.
-  if (AllInSet(forwardOp.inputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+  if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
-               no_grad_names)) {
     return NOP();
   }
 
   //  All output gradients of forwarding operator do not need to calculate.
   //  Then all input gradients cannot be computed at all, and we put them into
   //  `no_grad_names` set. Return an NOP.
-  if (AllInSet(forwardOp.outputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+  if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
-               no_grad_names)) {
     for (auto& name : forwardOp.inputs_) {
       // Mark all input is not need
-      no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+      no_grad_names.insert(name + kGradVarSuffix);
     }
     return NOP();
   }
@@ -134,9 +132,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
     for (std::string& grad_input : grad_op->inputs_) {
       if (no_grad_names.count(grad_input)) {
-        std::string prefix = grad_input.substr(
+        std::string prefix =
-            0, grad_input.size() - OperatorBase::GRAD_VAR_SUFFIX().size());
+            grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
-        grad_input = prefix + OperatorBase::ZERO_VAR_SUFFIX();
+        grad_input = prefix + kZeroVarSuffix;
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
@@ -147,7 +145,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
 
     for (std::string& grad_output : grad_op->outputs_) {
       if (no_grad_names.count(grad_output)) {
-        grad_output = OperatorBase::EMPTY_VAR_NAME();
+        grad_output = kEmptyVarName;
       }
     }
 
@@ -168,14 +166,14 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
-  no_grad_names.insert(OperatorBase::EMPTY_VAR_NAME() +
+  no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
-                       OperatorBase::GRAD_VAR_SUFFIX());
 
   for (auto& name : no_grad_vars) {
-    no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+    no_grad_names.insert(name + kGradVarSuffix);
   }
   size_t uid = 0;
   return BackwardRecursive(forwardOp, no_grad_names, uid);
 }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/backward_test.cc

@@ -78,14 +78,14 @@ class FcOp : public ops::NetOp {
                                {Output("mul_result")}, {}));
     auto b_name = Input("b");
     std::string before_act = "mul_result";
-    if (b_name != EMPTY_VAR_NAME()) {
+    if (b_name != kEmptyVarName) {
       AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name},
                                  {Output("add_result")}, {}));
       before_act = "add_result";
     } else {
       auto out_varname = Output("add_result");
-      if (out_varname != EMPTY_VAR_NAME()) {
+      if (out_varname != kEmptyVarName) {
-        this->Rename(out_varname, EMPTY_VAR_NAME());
+        this->Rename(out_varname, kEmptyVarName);
       }
     }
 
@@ -163,13 +163,12 @@ TEST(Backward, simple_op_grad) {
   ASSERT_NE(fwd, nullptr);
   auto gop = f::OpRegistry::CreateGradOp(*fwd);
   ASSERT_EQ(4UL, gop->inputs_.size());
-  ASSERT_EQ(f::OperatorBase::EMPTY_VAR_NAME(), gop->inputs_[0]);
+  ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
   ASSERT_EQ("rowwise_add_grad", gop->type_);
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[0]);
+  ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
-  ASSERT_EQ("b" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[1]);
+  ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
 
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+  ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
-            gop->Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()));
 }
 
 TEST(Backward, simple_op_not_need_grad) {
@@ -177,7 +176,7 @@ TEST(Backward, simple_op_not_need_grad) {
   ASSERT_NE(fwd, nullptr);
   auto gop = f::Backward(*fwd, {"X"});
   ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
-                      "X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+                      "X" + f::kGradVarSuffix),
             gop->outputs_.end());
 
   auto no_input_gop = f::Backward(*fwd, {"X", "b"});
@@ -210,9 +209,9 @@ TEST(Backward, net_fc_backward_normal) {
 }
 
 TEST(Backward, net_fc_backward_not_have_b) {
-  std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
+  std::shared_ptr<f::OperatorBase> fwd =
-      "fc", {"X", "w", f::OperatorBase::EMPTY_VAR_NAME()},
+      f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName},
-      {"mul_result", "add_result", "tmp"}, {});
+                              {"mul_result", "add_result", "tmp"}, {});
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
@@ -242,24 +241,21 @@ TEST(Backward, net_input_of_network_not_need_grad) {
 
   std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
       bwd_net->outputs_.begin(), bwd_net->outputs_.end());
-  all_output.erase(f::OperatorBase::EMPTY_VAR_NAME());
+  all_output.erase(f::kEmptyVarName);
 
   for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
-    ASSERT_NE(all_output.find(out + f::OperatorBase::GRAD_VAR_SUFFIX()),
+    ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end());
-              all_output.end());
   }
 
   // Not Generated X
-  ASSERT_EQ(all_output.find("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end());
-            all_output.end());
 
   ASSERT_EQ(2UL, bwd_net->ops_.size());
   ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
   auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
   ASSERT_EQ(3UL, first_fc_grad->ops_.size());
-  ASSERT_EQ(
+  ASSERT_EQ(f::kEmptyVarName,
-      f::OperatorBase::EMPTY_VAR_NAME(),
+            first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix));
-      first_fc_grad->ops_[2]->Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()));
 }
 
 TEST(Backward, net_shared_weight) {
@@ -311,17 +307,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
   ASSERT_EQ(1UL, fill_zero.inputs_.size());
   ASSERT_EQ("Z", fill_zero.inputs_[0]);
   ASSERT_EQ(1UL, fill_zero.outputs_.size());
-  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(), fill_zero.outputs_[0]);
+  ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
 
   auto &d_many_out = *net->ops_[1];
   ASSERT_EQ("many_output_op_grad", d_many_out.type_);
   ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
-  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(),
+  ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
-            d_many_out.Input("z" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
-  ASSERT_EQ("Y" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+  ASSERT_EQ("X" + f::kGradVarSuffix,
-            d_many_out.Input("y" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+            d_many_out.Output("x" + f::kGradVarSuffix));
-  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-            d_many_out.Output("x" + f::OperatorBase::GRAD_VAR_SUFFIX()));
 }
 
 TEST(Backward, op_part_of_input_are_not_need) {
@@ -331,12 +325,10 @@ TEST(Backward, op_part_of_input_are_not_need) {
   ASSERT_EQ(grad_mul.type_, "mul_grad");
   ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
   ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
-  ASSERT_EQ(grad_mul.Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  ASSERT_EQ(grad_mul.Output("A" + f::kGradVarSuffix), f::kEmptyVarName);
-            f::OperatorBase::EMPTY_VAR_NAME());
+  ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
-  ASSERT_EQ(grad_mul.Output("B" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
-            "b" + f::OperatorBase::GRAD_VAR_SUFFIX());
+            "out" + f::kGradVarSuffix);
-  ASSERT_EQ(grad_mul.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            "out" + f::OperatorBase::GRAD_VAR_SUFFIX());
   ASSERT_EQ(grad_mul.Input("A"), "a");
   ASSERT_EQ(grad_mul.Input("B"), "b");
   ASSERT_EQ(grad_mul.Input("Out"), "out");
@@ -368,23 +360,4 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL);
-
-  /*
-    EXPECT_EQ(grad_fc.Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-              f::OperatorBase::EMPTY_VAR_NAME());
-  EXPECT_EQ(grad_fc.Output("W" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-    "w3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Output("b" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-    "b3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Output("mul_result" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-  "mul_out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-
-  EXPECT_EQ(grad_fc.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-  "out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
-  EXPECT_EQ(grad_fc.Input("X"), "out2");
-  EXPECT_EQ(grad_fc.Input("W"), "w3");
-  EXPECT_EQ(grad_fc.Input("mul_result"), "mul_out3");
-  EXPECT_EQ(grad_fc.Input("add_result"), "tmp_out3");
-  EXPECT_EQ(grad_fc.Input("Out"), "out3");
-  */
 }

paddle/framework/ddim.h

@@ -25,18 +25,15 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-namespace {
-typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
-                       Dim<8>, Dim<9>>
-    DDimVar;
-}
-
 /**
  * \brief A dynamically sized dimension.
  *
  * The number of dimensions must be between [1, 9].
  */
 struct DDim {
+  typedef boost::variant<Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>, Dim<7>,
+                         Dim<8>, Dim<9>>
+      DDimVar;
   DDimVar var;
 
   DDim() : var(Dim<1>()) {}

paddle/framework/grad_op_builder.cc

@@ -56,8 +56,7 @@ static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
 
   for (const auto& arg : src_arg_list) {
     std::string src_name = arg.name();
-    std::string dst_name =
+    std::string dst_name = is_grad ? src_name + kGradVarSuffix : src_name;
-        is_grad ? src_name + OperatorBase::GRAD_VAR_SUFFIX() : src_name;
     (*dst_op->in_out_idxs_)[dst_name] = idx++;
     int src_arg_idx = src_op->in_out_idxs_->at(src_name);
     int src_begin =
@@ -65,10 +64,9 @@ static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
     int src_end = src_format == nullptr ? src_arg_idx + 1
                                         : src_format->at(src_arg_idx + 1);
     for (int i = src_begin; i < src_end; ++i) {
-      std::string s = is_grad ? src_inout[i] + OperatorBase::GRAD_VAR_SUFFIX()
+      std::string s =
-                              : arg.ignore_gradient()
+          is_grad ? src_inout[i] + kGradVarSuffix
-                                    ? OperatorBase::EMPTY_VAR_NAME()
+                  : (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]);
-                                    : src_inout[i];
       dst_inout.emplace_back(s);
     }
     if (dst_format != nullptr) {

paddle/framework/grad_op_builder.h

@@ -1,3 +1,17 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
 
 #include "paddle/framework/operator.h"

paddle/framework/grad_op_builder_test.cc

@@ -83,24 +83,21 @@ TEST(GradOpBuilder, MutiInOut) {
   EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
   EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
             std::vector<std::string>({"out2_1", "out2_2"}));
-  EXPECT_EQ(grad_test_op->Input("Out1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
-            "out1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+            "out1" + f::kGradVarSuffix);
-  EXPECT_EQ(
+  EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
-      grad_test_op->Inputs("Out2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            std::vector<std::string>(
-      std::vector<std::string>(
+                {"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
-          {"out2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-           "out2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
 
   ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(
+  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
-      grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
-      std::vector<std::string>({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                      "in2_2" + f::kGradVarSuffix,
-                                "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                      "in2_3" + f::kGradVarSuffix}));
-                                "in2_3" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+  EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
-  EXPECT_EQ(grad_test_op->Output("In3" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "in3" + f::kGradVarSuffix);
-            "in3" + f::OperatorBase::GRAD_VAR_SUFFIX());
 }
 
 TEST(GradOpBuilder, IOIgnoredInGradient) {
@@ -116,30 +113,25 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
   ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
   EXPECT_EQ(grad_test_op->Input("In1"), "in1");
   EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
-            std::vector<std::string>({f::OperatorBase::EMPTY_VAR_NAME(),
+            std::vector<std::string>({f::kEmptyVarName, f::kEmptyVarName}));
-                                      f::OperatorBase::EMPTY_VAR_NAME()}));
   EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
             std::vector<std::string>({"in3_1", "in3_2"}));
   EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
             std::vector<std::string>({"out1_1", "out1_2"}));
-  EXPECT_EQ(grad_test_op->Input("Out2"), f::OperatorBase::EMPTY_VAR_NAME());
+  EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
-  EXPECT_EQ(
+  EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix),
-      grad_test_op->Inputs("Out1_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            std::vector<std::string>(
-      std::vector<std::string>(
+                {"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
-          {"out1_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+  EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
-           "out1_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+            "out2" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Input("Out2" + f::OperatorBase::GRAD_VAR_SUFFIX()),
-            "out2" + f::OperatorBase::GRAD_VAR_SUFFIX());
 
   ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(
+  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
-      grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            std::vector<std::string>(
-      std::vector<std::string>({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                {"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
-                                "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+  EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
-  EXPECT_EQ(
+            std::vector<std::string>(
-      grad_test_op->Outputs("In3_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+                {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
-      std::vector<std::string>({"in3_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
-                                "in3_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
 }

paddle/framework/op_desc.proto

@@ -15,7 +15,7 @@ limitations under the License. */
 syntax = "proto2";
 package paddle.framework;
 
-import "attr_type.proto";
+import "attribute.proto";
 
 // AttrDesc is used to describe Attributes of an Operator. It contain's
 // name, type, and value of Attribute.

paddle/framework/op_proto.proto

@@ -22,7 +22,7 @@ limitations under the License. */
 syntax = "proto2";
 package paddle.framework;
 
-import "attr_type.proto";
+import "attribute.proto";
 
 // Attribute protocol message for 3rd-party language binding.
 // It will store the Op support what attribute and what type.

paddle/framework/op_registry.cc

@@ -14,37 +14,8 @@ limitations under the License. */
 
 #include <paddle/framework/op_registry.h>
 
-namespace paddle {
+#include <vector>
-namespace framework {
-
-template <>
-void AttrTypeHelper::SetAttrType<int>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::INT);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<float>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::FLOAT);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::string>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::STRING);
-}
 
-template <>
+namespace paddle {
-void AttrTypeHelper::SetAttrType<std::vector<int>>(AttrProto* attr) {
+namespace framework {}  // namespace framework
-  attr->set_type(paddle::framework::AttrType::INTS);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::vector<float>>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::FLOATS);
-}
-
-template <>
-void AttrTypeHelper::SetAttrType<std::vector<std::string>>(AttrProto* attr) {
-  attr->set_type(paddle::framework::AttrType::STRINGS);
-}
-}  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -19,7 +19,7 @@ limitations under the License. */
 #include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
-#include "paddle/framework/attr_checker.h"
+#include "paddle/framework/attribute.h"
 #include "paddle/framework/grad_op_builder.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/scope.h"
@@ -27,49 +27,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-// helper class to set attribute type
-struct AttrTypeHelper {
-  template <typename T>
-  static void SetAttrType(AttrProto* attr);
-
-  static Attribute GetAttrValue(const AttrDesc& attr_desc) {
-    switch (attr_desc.type()) {
-      case paddle::framework::AttrType::INT: {
-        return attr_desc.i();
-      }
-      case paddle::framework::AttrType::FLOAT: {
-        return attr_desc.f();
-      }
-      case paddle::framework::AttrType::STRING: {
-        return attr_desc.s();
-      }
-      case paddle::framework::AttrType::INTS: {
-        std::vector<int> val(attr_desc.ints_size());
-        for (int i = 0; i < attr_desc.ints_size(); ++i) {
-          val[i] = attr_desc.ints(i);
-        }
-        return val;
-      }
-      case paddle::framework::AttrType::FLOATS: {
-        std::vector<float> val(attr_desc.floats_size());
-        for (int i = 0; i < attr_desc.floats_size(); ++i) {
-          val[i] = attr_desc.floats(i);
-        }
-        return val;
-      }
-      case paddle::framework::AttrType::STRINGS: {
-        std::vector<std::string> val(attr_desc.strings_size());
-        for (int i = 0; i < attr_desc.strings_size(); ++i) {
-          val[i] = attr_desc.strings(i);
-        }
-        return val;
-      }
-    }
-    PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
-    return boost::blank();
-  }
-};
-
 // this class not only make proto but also init attribute checkers.
 class OpProtoAndCheckerMaker {
  public:
@@ -136,7 +93,7 @@ class OpProtoAndCheckerMaker {
     *attr->mutable_name() = name;
     *attr->mutable_comment() = comment;
     attr->set_generated(generated);
-    AttrTypeHelper::SetAttrType<T>(attr);
+    attr->set_type(AttrTypeID<T>());
     return op_checker_->AddAttrChecker<T>(name);
   }
 
@@ -297,7 +254,7 @@ class OpRegistry {
 
     AttributeMap attrs;
     for (auto& attr : op_desc.attrs()) {
-      attrs[attr.name()] = AttrTypeHelper::GetAttrValue(attr);
+      attrs[attr.name()] = GetAttrValue(attr);
     }
 
     return CreateOp(op_desc.type(), inputs, outputs, attrs);
@@ -314,7 +271,7 @@ class OpRegistry {
   static std::unordered_map<std::string, OpProto>& protos() {
     static std::unordered_map<std::string, OpProto> protos_;
     return protos_;
-  };
+  }
 
   static std::unordered_map<std::string, std::string>& grad_ops() {
     static std::unordered_map<std::string, std::string> grad_ops_;
@@ -336,12 +293,12 @@ class OpRegistry {
   static std::unordered_map<std::string, OpAttrChecker>& op_checkers() {
     static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
     return op_checkers_;
-  };
+  }
 
   static void GenerateTempVariableName(OperatorBase* op) {
     static std::atomic<size_t> gUniqId(0UL);
     for (auto& outname : op->outputs_) {
-      if (outname == OperatorBase::TMP_VAR_NAME()) {
+      if (outname == kTempVarName) {
         outname += op->type_;
         outname += "@";
         outname += std::to_string(gUniqId.fetch_add(1));
@@ -353,7 +310,7 @@ class OpRegistry {
 template <typename OpType, typename ProtoMakerType>
 class OpRegisterHelper {
  public:
-  OpRegisterHelper(const char* op_type) {
+  explicit OpRegisterHelper(const char* op_type) {
     OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
   }
 };

paddle/framework/operator.h

@@ -20,7 +20,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <vector>
 
-#include "paddle/framework/attr_checker.h"
+#include "paddle/framework/attribute.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/scope.h"
@@ -32,9 +32,29 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+/// If a variable is a empty variable, that name will be used.
+const std::string kEmptyVarName = "@EMPTY@";
+
+/// If a variable is a temporary variable, that name will be set in Python,
+/// but it will be convert to a unique name in scope after OpCreator.
+const std::string kTempVarName = "@TEMP@";
+
+/// If a variable's name has a certain suffix, it means that the
+/// variable is the gradient of another varibale.
+/// e.g. Variable "x@GRAD" is the gradient of varibale "x".
+const std::string kGradVarSuffix = "@GRAD";
+
+/// Variables with this suffix are supposed to be filled up with zeros.
+const std::string kZeroVarSuffix = "@ZERO";
+
+inline std::string GradVarName(const std::string& var_name) {
+  return var_name + kGradVarSuffix;
+}
+
 class OperatorBase;
 class InferShapeContext;
 class ExecutionContext;
+
 /**
  * OperatorBase has the basic element that Net will call to do computation.
  * Only CreateOperator from OpRegistry will new Operator directly. User
@@ -43,25 +63,6 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
-  /// If a variable is a empty variable, that name will be used.
-  static std::string EMPTY_VAR_NAME() { return "@EMPTY@"; }
-
-  /// If a variable is a temporary variable, that name will be set in Python,
-  /// but it will be convert to a unique name in scope after OpCreator.
-  static std::string TMP_VAR_NAME() { return "@TEMP@"; }
-
-  /// If a variable's name has a certain suffix, it means that the
-  /// variable is the gradient of another varibale.
-  /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
-  static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; }
-
-  static std::string GRAD_VAR_NAME(const std::string& name) {
-    return name + GRAD_VAR_SUFFIX();
-  }
-
-  /// Variables with this suffix are supposed to be filled up with zeros.
-  static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; }
-
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -284,7 +285,7 @@ class OperatorWithKernel : public OperatorBase {
     platform::Place place_;
 
     OpKernelKey() = default;
-    OpKernelKey(const platform::DeviceContext& dev_ctx) {
+    explicit OpKernelKey(const platform::DeviceContext& dev_ctx) {
       place_ = dev_ctx.GetPlace();
     }
 

paddle/framework/pybind.cc

@@ -105,7 +105,16 @@ PYBIND11_PLUGIN(core) {
       .def("set", PyCUDATensorSetFromArray<float>)
       .def("set", PyCUDATensorSetFromArray<int>)
 #endif
-      .def("shape", [](Tensor &self) { return vectorize(self.dims()); });
+      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
+      .def("set_float_element",
+           [](Tensor &self, size_t offset, float f) {
+             // TODO(yuyang18): Only support GPU now.
+             self.data<float>()[offset] = f;
+           })
+      .def("get_float_element", [](Tensor &self, size_t offset) -> float {
+        // TODO(yuyang18): Only support GPU now.
+        return self.data<float>()[offset];
+      });
 
   py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 
@@ -154,8 +163,8 @@ All parameter, weight, gradient are variables in Paddle.
   m.def_submodule(
        "var_names",
        "The module will return special predefined variable name in Paddle")
-      .def("empty", OperatorBase::EMPTY_VAR_NAME)
+      .def("empty", []() { return kEmptyVarName; })
-      .def("temp", OperatorBase::TMP_VAR_NAME);
+      .def("temp", []() { return kTempVarName; });
   // clang-format off
   py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
       .def_static("create",

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -967,8 +967,9 @@ void RecurrentGradientMachine::generateSequence() {
   size_t numSequences = getGenBatchSize();
 
   resizeBootFrame(numSequences);
-  // We create only two sub-network in generation for alternate use.
+  // We create only two sub-network in generation, one stores states of all
-  // Thus, we can reduce total memory of output_ in layer forward.
+  // layers in previous time step and the other storing the states at current
+  // time step.
   resizeOrCreateFrames(2);
 
   // outFrameLines_.size() > 1UL
@@ -1001,10 +1002,9 @@ void RecurrentGradientMachine::generateSequence() {
 
   // init outArg
   size_t resultNum = generator_.config.num_results_per_sample();
-  IVector::resizeOrCreate(
+  size_t maxGenWordCount =
-      generator_.outArg.ids,
+      generator_.config.max_num_frames() * numSequences * resultNum;
-      generator_.config.max_num_frames() * numSequences * resultNum,
+  IVector::resizeOrCreate(generator_.outArg.ids, maxGenWordCount, false);
-      false);
   if (resultNum > 1) {
     CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
     Matrix::resizeOrCreate(generator_.outArg.in,
@@ -1012,6 +1012,11 @@ void RecurrentGradientMachine::generateSequence() {
                            /* width */ resultNum,
                            false,
                            /* useGpu */ false);
+    Matrix::resizeOrCreate(generator_.outArg.value,
+                           /* height */ maxGenWordCount,
+                           /* width */ 1,
+                           false,
+                           /* useGpu */ false);
   }
   ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                 numSequences + 1,
@@ -1313,13 +1318,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
   starts[0] = 0;
   if (numResults > 1) {
     real* probs = generator_.outArg.in->getData();
+    real* idsProb = generator_.outArg.value->getData();
+    size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
         Path& path = finalPaths_[i][j];
-        generator_.ids.push_back(path.ids.size());  // sequence size
+        size_t genLen = path.ids.size();
+        generator_.ids.push_back(genLen);  // sequence size
         generator_.ids.insert(
             generator_.ids.end(), path.ids.begin(), path.ids.end());
         generator_.ids.push_back(-1);  // end of sequence
+
+        memcpy(idsProb + curPos, path.idsProb.data(), sizeof(real) * genLen);
+        curPos += genLen;
+        idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
 
         if (!j && dataArgsSize_) {

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -189,6 +189,11 @@ public:
      */
     std::vector<int> ids;
 
+    /**
+     * @brief idsProb, log probability of each generated words.
+     */
+    std::vector<real> idsProb;
+
     /**
      * @brief logProb, current probability of path.
      */
@@ -228,11 +233,13 @@ public:
      */
     Path(Path& old, int newId, real logProb, int machineId, int topIndex)
         : ids(old.ids),
+          idsProb(old.idsProb),
           logProb(old.logProb + logProb),
           machineId(machineId),
           topIndex(topIndex),
           seqId(old.seqId) {
       ids.push_back(newId);
+      idsProb.push_back(logProb);
       if (!old.probHistory.empty()) {
         this->probHistory = old.probHistory;
         // probHistory store current prob, not sum
@@ -411,8 +418,9 @@ protected:
 
   struct Generator {
     GeneratorConfig config;
-    std::vector<int> ids;  // store generated sequences
+    std::vector<int> ids;       // store generated sequences
-    Argument outArg;       // final output argument
+    std::vector<real> idsProb;  // log probability of each generated word
+    Argument outArg;            // final output argument
   };
   bool generating_;
   Generator generator_;

paddle/gserver/tests/CMakeLists.txt

@@ -1,5 +1,10 @@
 # gserver pacakge unittests
 
+file(GLOB_RECURSE GSERVER_HEADER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.h")
+file(GLOB_RECURSE GSERVER_SOURCES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.cpp")
+add_style_check_target(paddle_gserver ${GSERVER_SOURCES})
+add_style_check_target(paddle_gserver ${GSERVER_HEADER})
+
 ################### test_ProtoDataProvider ############
 add_unittest_without_exec(test_ProtoDataProvider
     test_ProtoDataProvider.cpp)
@@ -50,7 +55,7 @@ add_unittest_without_exec(test_DetectionOutput
     test_DetectionOutput.cpp
     LayerGradUtil.cpp)
 
-add_test(NAME test_DetectionOutput 
+add_test(NAME test_DetectionOutput
     COMMAND test_DetectionOutput)
 ################# test_ConvUnify #######################
 add_unittest_without_exec(test_ConvUnify

paddle/gserver/tests/LayerGradUtil.cpp

@@ -400,7 +400,6 @@ void initDataLayer(TestConfig testConf,
         const std::vector<int>& labelSeqStartPositions =
             testConf.inputDefs[i].labelSeqStartPositions;
         if (labelSeqStartPositions.size() != 0) {
-          CHECK(!sequenceStartPositions);
           CHECK_GE(static_cast<int>(labelSeqStartPositions.size()), 2);
 
           sequenceStartPositions =
@@ -410,6 +409,19 @@ void initDataLayer(TestConfig testConf,
                                            useGpu);
           data.sequenceStartPositions = sequenceStartPositions;
         }
+
+        const std::vector<int>& labelSubSeqStartPositions =
+            testConf.inputDefs[i].labelSubSeqStartPositions;
+        if (labelSubSeqStartPositions.size() != 0) {
+          CHECK_GE(static_cast<int>(labelSubSeqStartPositions.size()), 2);
+
+          subSequenceStartPositions =
+              ICpuGpuVector::create(labelSubSeqStartPositions.size(), useGpu);
+          subSequenceStartPositions->copyFrom(labelSubSeqStartPositions.data(),
+                                              labelSubSeqStartPositions.size(),
+                                              useGpu);
+          data.subSequenceStartPositions = subSequenceStartPositions;
+        }
         break;
       }
       default:

paddle/gserver/tests/LayerGradUtil.h

@@ -67,6 +67,7 @@ struct InputDef {
   bool isStatic;
   std::vector<int> labelInitValue;
   std::vector<int> labelSeqStartPositions;
+  std::vector<int> labelSubSeqStartPositions;
   MatrixPtr selfDefinedData;
 
   InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) {
@@ -81,8 +82,10 @@ struct InputDef {
   InputDef(InputType type,
            string nameIn,
            MatrixPtr selfDefinedData,
-           std::vector<int> selfDefinedSeqStartPos = {})
+           std::vector<int> selfDefinedSeqStartPos = {},
+           std::vector<int> selfDefinedSubSeqStartPos = {})
       : labelSeqStartPositions(selfDefinedSeqStartPos),
+        labelSubSeqStartPositions(selfDefinedSubSeqStartPos),
         selfDefinedData(selfDefinedData) {
     inputType = type;
     name = nameIn;